An AI use clause for client contracts

The standard IP clause says the agency assigns all right, title and interest in the deliverables to the client on payment. It worked for decades because it was drafted for a world where somebody definitely owned something.

Run it over generated work and it quietly promises something no agency can deliver. In the US, human authorship is required and purely AI-generated material is unregistrable. The Supreme Court denied certiorari in Thaler v. Perlmutter on 2 March 2026, leaving that line where the D.C. Circuit put it. You cannot assign copyright that does not exist. The clause is not wrong exactly; it is empty in the places that matter, and both sides find out which places those are at the worst possible moment.

What follows is the structure of an AI use clause and drafting language to adapt. It is a starting point for a conversation with counsel, not legal advice.

Start with a definition, because "AI" is not one

Half the disputes here are vocabulary. A client who says "no AI" usually means no synthetic humans and no generated brand imagery, not no noise reduction, no auto-transcription and no generative fill on a dust spot. If the agreement does not distinguish, you find out during the first review round.

AI-Assisted Material means any Deliverable or component of a Deliverable produced in whole or in part by a generative model, including image, video, audio, voice and music generation, and generative editing operations that create new content not present in the source. It does not include analytical or corrective processing such as transcription, alignment, noise reduction, upscaling, colour correction or object removal that does not synthesise new subject matter.

Draw the line wherever the relationship needs it. Draw it somewhere.

Clause 1: disclosure of involvement, agency to client

This is a representation, not a marketing decision. The client is entitled to know what they are buying, and the agency is protected by having said so in writing.

Supplier will identify, at delivery, which Deliverables constitute AI-Assisted Material, the generative models used, and whether any real person's likeness or voice was synthesised. Supplier will maintain a record of this information for the duration of the licence term and provide it on reasonable request.

That record makes Clause 3 actionable and gives Clause 2 its evidence base.

Clause 2: ownership, with its ceiling stated

The honest clause separates three things the standard one smears together: the files, the permitted uses, and copyright.

Supplier assigns to Client all right, title and interest that Supplier holds in the Deliverables. Client acknowledges that copyright protection in AI-Assisted Material depends on human authorship and may be limited or unavailable in some jurisdictions, and that Supplier makes no representation that any Deliverable is protectable by copyright. Supplier grants Client unrestricted use of the delivered files for the purposes set out in Schedule [X], irrespective of their copyright status.

That last sentence keeps the deal working. What the client operationally wants is to run the asset everywhere, forever, without anyone stopping them. That is a licence question, and it survives whatever copyright says. Two hard limits sit behind it.

A provider can only assign what it holds. OpenAI's terms assign its right, title and interest in Output to the user and permit commercial use. Midjourney's paid subscribers own their assets subject to revenue-threshold tiering and a broad licence back, with IP warranties disclaimed on consumer plans. None of that creates copyright where the human-authorship test fails. "You own the output" and "the output is protectable" are different sentences.

Jurisdiction changes the answer. The EU has no harmonised AI-authorship rule; the CJEU's "author's own intellectual creation" standard governs and is untested against generative AI. In the UK, CDPA s.9(3) on computer-generated works remains on the books but its application here is contested, and the government's March 2026 report on copyright and AI abandoned the proposed broad text-and-data-mining exception. A clause assuming one answer travels badly.

If the client needs enforceability, add a step rather than a promise.

Where Client requires a Deliverable to be protectable, the parties will agree in advance the human authorship contribution to be applied, and Supplier will record the creative decisions constituting that contribution.

Human-authored selection, arrangement and modification are what is protectable, and AI-assisted works are registrable in the US with disclosure of the AI material. The record has to be made while the work happens. Our copyright and safety guide covers the distinction.

Clause 3: who carries the disclosure duty

This is the clause most agreements are missing, and the one with a hard date. EU AI Act Article 50 applies from 2 August 2026. Its deployer duty, labelling deepfakes and disclosing to the viewer on first exposure rather than in metadata, attaches to whoever puts the content out. In an agency relationship that is very often the client, who holds the media plan and the publishing accounts. New York's synthetic performer disclosure, effective 9 June 2026, hits the advertiser whose ad reaches New York consumers. In the UK, the ASA's position is that the advertiser is fully responsible for AI-produced content.

The duty lands on the party that publishes. Say so, and say what the other party owes.

Client is responsible for compliance with disclosure, labelling and transparency obligations applicable to publication of the Deliverables in each territory of use. Supplier will provide the information in Clause 1, deliver assets with any provenance metadata intact where technically feasible, and implement specific disclosure treatments requested by Client at the point of production. Client will notify Supplier of the territories of use before production begins.

That last sentence does more work than it looks. Territory decides which regime applies and is routinely settled after the creative is finished; pulling it forward into an obligation is the cheapest fix available. The Article 50 breakdown for sponsored content covers where the deployer line falls. Note that "where technically feasible" is also doing real work: a C2PA manifest is a hard binding to the file, and re-encoding or non-C2PA-aware editing destroys it.

Clause 4: indemnity that matches who controls the risk

The default position, where the supplier indemnifies the client for third-party IP claims arising from deliverables, is uncomfortable for generated work: the supplier cannot fully control what a model produces, and holds no upstream indemnity that flows through. Provider indemnities are narrower than the marketing suggests. Adobe offers IP indemnification on qualifying commercial and enterprise Firefly plans, with contract-specific caps and conditions. Google, Microsoft and Anthropic generally offer copyright indemnities at enterprise or API tiers, conditioned on using safety filters, not on consumer tiers. An indemnity your vendor gives you is not automatically one you can pass on.

Allocate by control instead.

Supplier indemnifies Client against claims arising from Supplier's breach of Clause 1, from Supplier's failure to obtain consents it undertook to obtain, and from Supplier's use of source material it warranted it had the right to use. Client indemnifies Supplier against claims arising from Client-supplied material, from Client's use of the Deliverables outside the scope in Schedule [X], and from Client's failure to apply disclosures under Clause 3. Neither party indemnifies the other against claims arising solely from the training data or inherent output characteristics of a third-party generative model, which the parties acknowledge is an unsettled area of law.

The final carve-out is the honest one. Litigation over outputs depicting protected material is live and unresolved, and neither an agency nor a client can underwrite it. Say so rather than allocating it by accident to whoever signed the more standard template. Agencies working under a white-label arrangement should check the carve-out survives being passed down a chain.

Clause 5: likeness and voice get their own paragraph

A general AI clause does not cover a replica, and a general consent does not either. California's AB 2602, effective 1 January 2025, makes digital-replica clauses in performer contracts unenforceable without a specific use description and professional representation. That is a drafting constraint, not background colour: a broad "we may create digital replicas" clause may simply not work.

No Deliverable will synthesise the likeness or voice of an identifiable individual except as expressly listed in Schedule [Y], and only where the consent obtained describes the specific use, medium, territory and term. Voice consent is separate from likeness consent. Any model, embedding or replica artefact created for this purpose will be deleted at the end of the term unless Schedule [Y] provides otherwise.

The deletion sentence is what performers and their representatives care about most; offering it unprompted materially improves the negotiation.

Making it operational

Three habits keep the clause live rather than decorative.

1. Put the scope in a schedule, not the body. Territory, platforms, term, paid versus organic. The usage rights explainer and the detail in usage rights in creator contracts set the right granularity, and tiered rate cards show how to price the difference.
2. Connect it to the revision policy. "Regenerate it without the AI" is a scope event, and a revision policy that stops scope creep should name it.
3. Keep one version. Agencies negotiating this per-client end up with incompatible positions across a book of business. Our notes for agencies cover standardising.

FAQ

Should I just promise full copyright and hope?

No, and not for ethical reasons alone. A representation you cannot support is what a client points at when something goes wrong, and it converts an unsettled legal question into a straightforward breach claim against you. Stating the limit is protective.

Does a no-training representation belong in here?

It belongs if either party cares, and clients increasingly do. The relevant question is what your tooling actually does with submitted material, which varies by vendor and by tier. Confirm it against your own agreements before you represent it, because that is a factual statement about your supply chain rather than a legal position.

Who holds the disclosure obligation if we also buy the media?

Whoever publishes. If the agency runs the accounts and places the buy, the agency is putting the content out and the duty follows. Track the actual operating model rather than the default template, and revisit the clause when that model changes.